Agent Harness

What it does

Takes a single goal and compiles it into a verifiable task plan, then drives that plan through a durable, state-file-based loop until a legitimate close.

• goal_compiler.py: turns a goal into plan.json; vague goals are refused with exit 3 plus forcing questions.
• loop_controller.py: manages init → next → record → verify → close. Checks are run by the controller itself via subprocess, so the agent never gets to declare its own work verified.
• Budgets are terminal: default 3 attempts per task and 12 loop iterations; exhaustion escalates to a human instead of faking success.
• close is refused (exit 4) while any task is unverified; only a human may waive.
• State lives in .agent-harness/state.json, so a fresh session can resume from plan + state alone.

Who it's for

• Anyone tired of agents claiming "done" when nothing actually passes.
• Teams running long goals across multiple sessions who need reproducibility.
• Teams wrapping domain skill folders (engineering, marketing, finance…) into self-verifying loops.
• Engineers who want reward hacking (editing the gate to pass it) structurally blocked.

Examples

1. Payments SLO design: compile "audit the payments service and design an SLO with an error budget" against the engineering manifest; each task only closes when its check command exits 0.
2. Marketing loop: build a harness from the marketing domain manifest to run research → copy → checklist verification, escalating to a named reviewer after 3 failed attempts.
3. CI manifest check: regenerate manifests with harness_manifest_builder.py --no-timestamp and fail CI on any diff, keeping harnesses in sync as skills change.

Install the skill I found on Claude Skill Mart.
Copy the .gemini/skills/agent-harness folder from the GitHub repo alirezarezvani/claude-skills into my ~/.claude/skills/agent-harness/.
When it's done, tell me in one line what this skill can do.